Closed NixGD closed 2 years ago
The maths in the code is correct. Evidence:
During the saliency experiments, gradients were calculated wrt PC-K at timestep 11. Correspondingly, now that we're projecting the gradients of hx onto the PCs, the gradients should most strongly project onto PC-K at t=11. This is indeed what we see:
Also, import the new data set.